hongshaorou/Tianshou
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

0. code refactor, try to merge Go and GoEnv

Browse Source
This commit is contained in:
Dong Yan 2017-12-16 23:29:11 +08:00
parent 01c0c2483a
commit 31199c7d0d
5 changed files with 57 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
AlphaGo/game.py
View File

@ -31,14 +31,14 @@ class Game:
self.komi = komi self.komi = komi
self.board = [utils.EMPTY] * (self.size * self.size) self.board = [utils.EMPTY] * (self.size * self.size)
self.history = [] self.history = []
self.past = deque(maxlen=8) self.latest_boards = deque(maxlen=8)
for _ in range(8): for _ in range(8):
self.past.append(self.board) self.latest_boards.append(self.board)
self.executor = go.Go(game=self) self.executor = go.Go(game=self)
#self.strategy = strategy(checkpoint_path) #self.strategy = strategy(checkpoint_path)
self.simulator = strategy.GoEnv() self.simulator = strategy.GoEnv(game=self)
self.net = network_small.Network() self.net = network_small.Network()
self.sess = self.net.forward(checkpoint_path) self.sess = self.net.forward(checkpoint_path)
self.evaluator = lambda state: self.sess.run([tf.nn.softmax(self.net.p), self.net.v], self.evaluator = lambda state: self.sess.run([tf.nn.softmax(self.net.p), self.net.v],
@ -57,7 +57,7 @@ class Game:
self.board = [utils.EMPTY] * (self.size * self.size) self.board = [utils.EMPTY] * (self.size * self.size)
self.history = [] self.history = []
for _ in range(8): for _ in range(8):
self.past.append(self.board) self.latest_boards.append(self.board)
def set_size(self, n): def set_size(self, n):
self.size = n self.size = n
@ -66,29 +66,29 @@ class Game:
def set_komi(self, k): def set_komi(self, k):
self.komi = k self.komi = k
def data_process(self, history, color): def generate_nn_input(self, history, color):
state = np.zeros([1, self.simulator.size, self.simulator.size, 17]) state = np.zeros([1, self.size, self.size, 17])
for i in range(8): for i in range(8):
state[0, :, :, i] = np.array(np.array(history[i]) == np.ones(self.simulator.size ** 2)).reshape(self.simulator.size, self.simulator.size) state[0, :, :, i] = np.array(np.array(history[i]) == np.ones(self.size ** 2)).reshape(self.size, self.size)
state[0, :, :, i + 8] = np.array(np.array(history[i]) == -np.ones(self.simulator.size ** 2)).reshape(self.simulator.size, self.simulator.size) state[0, :, :, i + 8] = np.array(np.array(history[i]) == -np.ones(self.size ** 2)).reshape(self.size, self.size)
if color == utils.BLACK: if color == utils.BLACK:
state[0, :, :, 16] = np.ones([self.simulator.size, self.simulator.size]) state[0, :, :, 16] = np.ones([self.size, self.size])
if color == utils.WHITE: if color == utils.WHITE:
state[0, :, :, 16] = np.zeros([self.simulator.size, self.simulator.size]) state[0, :, :, 16] = np.zeros([self.size, self.size])
return state return state
def strategy_gen_move(self, history, color): def strategy_gen_move(self, latest_boards, color):
self.simulator.history = copy.copy(history) self.simulator.latest_boards = copy.copy(latest_boards)
self.simulator.board = copy.copy(history[-1]) self.simulator.board = copy.copy(latest_boards[-1])
state = self.data_process(self.simulator.history, color) nn_input = self.generate_nn_input(self.simulator.latest_boards, color)
mcts = MCTS(self.simulator, self.evaluator, state, self.simulator.size ** 2 + 1, inverse=True, max_step=10) mcts = MCTS(self.simulator, self.evaluator, nn_input, self.size ** 2 + 1, inverse=True, max_step=1)
temp = 1 temp = 1
prob = mcts.root.N ** temp / np.sum(mcts.root.N ** temp) prob = mcts.root.N ** temp / np.sum(mcts.root.N ** temp)
choice = np.random.choice(self.simulator.size ** 2 + 1, 1, p=prob).tolist()[0] choice = np.random.choice(self.size ** 2 + 1, 1, p=prob).tolist()[0]
if choice == self.simulator.size ** 2: if choice == self.size ** 2:
move = utils.PASS move = utils.PASS
else: else:
move = (choice % self.simulator.size + 1, choice / self.simulator.size + 1) move = (choice % self.size + 1, choice / self.size + 1)
return move, prob return move, prob
def do_move(self, color, vertex): def do_move(self, color, vertex):
@ -100,7 +100,7 @@ class Game:
def gen_move(self, color): def gen_move(self, color):
# move = self.strategy.gen_move(color) # move = self.strategy.gen_move(color)
# return move # return move
move, self.prob = self.strategy_gen_move(self.past, color) move, self.prob = self.strategy_gen_move(self.latest_boards, color)
self.do_move(color, move) self.do_move(color, move)
return move return move
@ -127,3 +127,6 @@ class Game:
if __name__ == "__main__": if __name__ == "__main__":
g = Game() g = Game()
g.show_board() g.show_board()
#file = open("debug.txt", "a")
#file.write("mcts check\n")
#file.close()

2
AlphaGo/go.py
View File

@ -135,7 +135,7 @@ class Go:
self.game.board[self.game._flatten(vertex)] = color self.game.board[self.game._flatten(vertex)] = color
self._process_board(color, vertex) self._process_board(color, vertex)
self.game.history.append(copy.copy(self.game.board)) self.game.history.append(copy.copy(self.game.board))
self.game.past.append(copy.copy(self.game.board)) self.game.latest_boards.append(copy.copy(self.game.board))
return True return True
def _find_empty(self): def _find_empty(self):

4
AlphaGo/play.py
View File

@ -87,8 +87,8 @@ if __name__ == '__main__':
score = player[turn].run_cmd(str(num) + ' get_score') score = player[turn].run_cmd(str(num) + ' get_score')
print "Finished : ", score.split(" ")[1] print "Finished : ", score.split(" ")[1]
player[0].run_cmd(str(num) + ' clear_board') player[0].run_cmd(str(num) + ' clear_board')
player[1].run_cmd(str(num) + ' clear_board') player[1].run_cmd(str(num) + ' clear_board')
game_num += 1 game_num += 1
subprocess.call(["kill", "-9", str(agent_v0.pid)]) subprocess.call(["kill", "-9", str(agent_v0.pid)])
subprocess.call(["kill", "-9", str(agent_v1.pid)]) subprocess.call(["kill", "-9", str(agent_v1.pid)])

57
AlphaGo/strategy.py
View File

@ -14,15 +14,14 @@ DELTA = [[1, 0], [-1, 0], [0, -1], [0, 1]]
CORNER_OFFSET = [[-1, -1], [-1, 1], [1, 1], [1, -1]] CORNER_OFFSET = [[-1, -1], [-1, 1], [1, 1], [1, -1]]
class GoEnv: class GoEnv:
def __init__(self, size=9, komi=6.5): def __init__(self, **kwargs):
self.size = size self.game = kwargs['game']
self.komi = komi self.board = [utils.EMPTY] * (self.game.size * self.game.size)
self.board = [utils.EMPTY] * (self.size * self.size) self.latest_boards = deque(maxlen=8)
self.history = deque(maxlen=8)
def _flatten(self, vertex): def _flatten(self, vertex):
x, y = vertex x, y = vertex
return (x - 1) * self.size + (y - 1) return (x - 1) * self.game.size + (y - 1)
def _bfs(self, vertex, color, block, status, alive_break): def _bfs(self, vertex, color, block, status, alive_break):
block.append(vertex) block.append(vertex)
@ -35,7 +34,7 @@ class GoEnv:
def _find_block(self, vertex, alive_break=False): def _find_block(self, vertex, alive_break=False):
block = [] block = []
status = [False] * (self.size * self.size) status = [False] * (self.game.size * self.game.size)
color = self.board[self._flatten(vertex)] color = self.board[self._flatten(vertex)]
self._bfs(vertex, color, block, status, alive_break) self._bfs(vertex, color, block, status, alive_break)
@ -73,7 +72,7 @@ class GoEnv:
_board = copy.copy(self.board) _board = copy.copy(self.board)
self.board[self._flatten(vertex)] = color self.board[self._flatten(vertex)] = color
self._process_board(color, vertex) self._process_board(color, vertex)
if self.board in self.history: if self.board in self.latest_boards:
res = True res = True
else: else:
res = False res = False
@ -83,8 +82,8 @@ class GoEnv:
def _in_board(self, vertex): def _in_board(self, vertex):
x, y = vertex x, y = vertex
if x < 1 or x > self.size: return False if x < 1 or x > self.game.size: return False
if y < 1 or y > self.size: return False if y < 1 or y > self.game.size: return False
return True return True
def _neighbor(self, vertex): def _neighbor(self, vertex):
@ -151,21 +150,28 @@ class GoEnv:
# print "many opponents, fake eye" # print "many opponents, fake eye"
return False return False
# def is_valid(self, color, vertex): def knowledge_prunning(self, color, vertex):
def is_valid(self, state, action): ### check if it is an eye of yourself
### assumptions : notice that this judgement requires that the state is an endgame
if self._is_eye(color, vertex):
return False
return True
def simulate_is_valid(self, state, action):
# state is the play board, the shape is [1, 9, 9, 17] # state is the play board, the shape is [1, 9, 9, 17]
if action == self.size * self.size: if action == self.game.size * self.game.size:
vertex = (0, 0) vertex = (0, 0)
else: else:
vertex = (action / self.size + 1, action % self.size + 1) vertex = (action / self.game.size + 1, action % self.game.size + 1)
if state[0, 0, 0, -1] == utils.BLACK: if state[0, 0, 0, -1] == utils.BLACK:
color = utils.BLACK color = utils.BLACK
else: else:
color = utils.WHITE color = utils.WHITE
self.history.clear() self.latest_boards.clear()
for i in range(8): for i in range(8):
self.history.append((state[:, :, :, i] - state[:, :, :, i + 8]).reshape(-1).tolist()) self.latest_boards.append((state[:, :, :, i] - state[:, :, :, i + 8]).reshape(-1).tolist())
self.board = copy.copy(self.history[-1]) self.board = copy.copy(self.latest_boards[-1])
### in board ### in board
if not self._in_board(vertex): if not self._in_board(vertex):
return False return False
@ -180,12 +186,11 @@ class GoEnv:
if not self._is_qi(color, vertex): if not self._is_qi(color, vertex):
return False return False
### check if it is an eye of yourself ### forbid global isomorphous
### assumptions : notice that this judgement requires that the state is an endgame if self._check_global_isomorphous(color, vertex):
if self._is_eye(color, vertex):
return False return False
if self._check_global_isomorphous(color, vertex): if not self.knowledge_prunning(color, vertex):
return False return False
return True return True
@ -206,17 +211,17 @@ class GoEnv:
color = utils.BLACK color = utils.BLACK
else: else:
color = utils.WHITE color = utils.WHITE
if action == self.size ** 2: if action == self.game.size ** 2:
vertex = utils.PASS vertex = utils.PASS
else: else:
vertex = (action % self.size + 1, action / self.size + 1) vertex = (action % self.game.size + 1, action / self.game.size + 1)
# print(vertex) # print(vertex)
# print(self.board) # print(self.board)
self.board = (state[:, :, :, 7] - state[:, :, :, 15]).reshape(-1).tolist() self.board = (state[:, :, :, 7] - state[:, :, :, 15]).reshape(-1).tolist()
self.do_move(color, vertex) self.do_move(color, vertex)
new_state = np.concatenate( new_state = np.concatenate(
[state[:, :, :, 1:8], (np.array(self.board) == utils.BLACK).reshape(1, self.size, self.size, 1), [state[:, :, :, 1:8], (np.array(self.board) == utils.BLACK).reshape(1, self.game.size, self.game.size, 1),
state[:, :, :, 9:16], (np.array(self.board) == utils.WHITE).reshape(1, self.size, self.size, 1), state[:, :, :, 9:16], (np.array(self.board) == utils.WHITE).reshape(1, self.game.size, self.game.size, 1),
np.array(1 - state[:, :, :, -1]).reshape(1, self.size, self.size, 1)], np.array(1 - state[:, :, :, -1]).reshape(1, self.game.size, self.game.size, 1)],
axis=3) axis=3)
return new_state, 0 return new_state, 0

2
tianshou/core/mcts/mcts.py
View File

@ -75,7 +75,7 @@ class UCTNode(MCTSNode):
start_time = time.time() start_time = time.time()
self.mask = [] self.mask = []
for act in range(self.action_num - 1): for act in range(self.action_num - 1):
if not simulator.is_valid(self.state, act): if not simulator.simulate_is_valid(self.state, act):
self.mask.append(act) self.mask.append(act)
self.ucb[act] = -float("Inf") self.ucb[act] = -float("Inf")
else: else: